In a nutshell

We propose several explanation methodologies building on Information Bottleneck Attribution (IBA):

• Our main contribution is the Inverse IBA saliency (attribution) method which identifies any region that has predictive information for the model.
• We also propose Regression IBA for explaining regression models. Using Regression IBA we observe that a model trained on cumulative COVID-19 severity score labels implicitly learns the severity of different X-ray regions.
• Finally, we propose Multi-layer IBA to generate higher resolution and detailed attribution/saliency maps.

Resources

View the paper on arXiv (The camera-ready version will appear in the proceedings of MICCAI 2021.)

Check the Code on GitHub

Citation

Please cite the work using the below BibTeX (also available on the Open Access link above)

@misc{khakzar2021explaining,
      title={Explaining COVID-19 and Thoracic Pathology Model Predictions by Identifying Informative Input Features}, 
      author={Ashkan Khakzar and Yang Zhang and Wejdene Mansour and Yuezhi Cai and Yawei Li and Yucheng Zhang and Seong Tae Kim and Nassir Navab},
      year={2021},
      eprint={2104.00411},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}

Contact

For inquiries and feedback please contact Ashkan Khakzar (ashkan.khakzar@tum.de). We would be happy to help and we appreciate your feedback.